Real-time Condition Monitoring of filling machines with vibration analysis and Edge AI

This paper presents an embedded sensor system, which allows Condition Monitoring (CM) on filling machines. An auger filler is used as an example, which doses food into bags. Due to a mechanical defect, metal shavings can occur during the operation of an auger filler. Those metal shavings can contaminate the dosed food and cause huge costs due to possible recall action and loss of credibility. The presented CM solution should detect such events by utilizing vibration data. We employ feature extraction and machine learning algorithms to classify the condition of the auger filler. The resulting sensor system consists of an Artificial Neural Network (ANN) for the CM. The feature extraction and the ANN were successfully ported to a microcontroller, allowing retrofitting to existing auger fillers and thus representing an application of Edge AI. On our test set, we achieve a classification accuracy of 100% for the dosing product milk powder, which is the main application of this system. With the ANN for all available dosing products, an accuracy of 96.6% is achieved. Furthermore, the system can monitor each dosing cycle of the auger filler, as the execution time of a condition prediction is lower (< 200 ms) than the cycle time of the auger filler (800 ms). Moreover, it can monitor the auger filler without additional settings for different configurations (e.g., dispensed food and cycle rate), eliminating the need to make individual adjustments, as is the case with systems currently in use.